Cross-Lingual Transfer Learning for Named Entity Recognition in Low-Resource Languages

This topic investigates the effectiveness of cross-lingual transfer learning for Named Entity Recognition (NER) in low-resource languages. The research focuses on pre-training models on high-resource languages and fine-tuning them on limited annotated datasets of low-resource languages. Special attention is given to addressing the challenges of morphological complexity and data scarcity in these languages. The study builds upon existing research, such as Wu and Dredze (2019), which demonstrated the effectiveness of mBERT in transferring knowledge to low-resource languages for NER tasks. Exploring techniques used in other sequence labeling tasks, such as LOREM, for Named Entity Recognition (NER).

Key research objectives include:

• Investigating state-of-the-art methods in cross-lingual NER
• Developing novel approaches for knowledge transfer from high-resource to low-resource languages
• Exploring innovative data augmentation techniques to generate labeled data for low-resource languages
• Implementing and evaluating transfer learning approaches to surpass existing benchmarks
• Investigating unsupervised and semi-supervised methods to improve training data availability

Research Gaps

Limited exploration of unsupervised methods for cross-lingual NER: Most existing research on cross-lingual NER has focused on supervised methods, which require labeled data in the source language. There is a gap in understanding how unsupervised methods can be effectively used for cross-lingual NER in low-resource language

Tasks

• Investigate the state-of-the-art in cross-lingual NER
• Develop an approach for transferring knowledge from a model trained on high resource language to low resource language.
• Data augmentation techniques for generating labeled data for low resource languages

Prerequisites

• Basic NLP concepts
• Proficiency with Python Programming
• Deep learning and Transfer learning concepts